a little cache for web page translation?

Some normal pages take up to 30s to translate with the giellatekno pairs.

Maybe we should try having a little
`{ pair : { url : translationOutput } }`
? Just for the slashdot effect where one page gets requested again and again.

Seeing as caching is Hard, we could keep it simple and just empty it when sys.getsizeof > max_url_cache. But sys.getsizeof doesn't recursively follow the size of the object; also, news sites etc. change, so I added a stupid simple "empty if older than two hours" rule for the whole cache object.

TODO:

- [ ] have a time stamp per url instead
- [ ] maybe don't save to cache if we already have N pages cached for that pair?
- [ ] fix it for py32 https://travis-ci.org/goavki/apertium-apy/jobs/197297317